What Work Does Generative AI Do? -- by Alexander Bick, Adam Blandin, David J. Deming, Tyler R. Schumacher

TEXT START: We measure how workers use genAI for their jobs in a nationally representative survey linking genAI adoption to detailed occupations and tasks.

The Dissection

This is an instrumentation paper, not a theory of labor-market survival. It builds task- and occupation-level adoption indexes, exposes weaknesses in crude exposure scores and platform-chat logs, and documents widespread but shallow use. Its real move is to shift attention from whether AI can perform work to who uses it now and where. Useful measurement; no systemic diagnosis.

The Core Fallacy

The implicit error is treating present adoption as the decisive variable. Under DT mechanics, current adoption is a lag indicator shaped by diffusion friction, worker choice, and organizational policy—not the long-run cost-and-performance boundary. Fewer than half of workers adopting can coexist with employers producing the same output with fewer people once AI-capital becomes competitively mandatory.

The paper measures assistance, not substitution, wage compression, ownership of the gains, or the collapse of productive participation. “Widespread but shallow” is not evidence of stability. It is the runway before competitive pressure deepens adoption.

Hidden Assumptions

• Adoption remains voluntary and incremental rather than imposed by competition.
• Occupations and tasks remain stable categories after AI reorganizes workflows.
• Survey-reported use adequately captures economic impact, including employer-side automation and workflow redesign.
• Worker heterogeneity matters more than ownership and control of AI capital.
• Variation among similar workers signals durable differentiation rather than temporary diffusion variance.
• A current snapshot is treated as central without establishing that it is a terminal equilibrium.

Social Function

Classification: partial truth with a transition-management function. The paper is not pure copium; its measurement distinctions are legitimate and necessary. But it domesticates a structural shock into indexes, surveys, and adopter heterogeneity, postponing the decisive question: what preserves mass bargaining power when equivalent output requires fewer humans?

Used carelessly, it becomes an ideological anesthetic—better measurement of diffusion mistaken for evidence against displacement.

The Verdict

The paper improves the map of current genAI use but does not challenge DT. It provides no evidence against durable AI superiority, coordination failure, or majority exclusion from economically necessary labor. Its central finding describes a lag phase, not survival. It records the corpse’s remaining reflexes; it does not demonstrate regeneration.